Estimation of HIV prevalence at the ZIP code-level in Atlanta, Georgia: Bayesian prediction modeling using passive surveillance data and social determinants of disease spreading

Objective

This study aims to predict the number of undiagnosed HIV cases at the ZIP Code-level in Atlanta, Georgia, based on publicly available information.

Study design

Statistical modeling.

Methods

We fitted a Bayesian hierarchical Binomial model to county-level estimates of the passive-surveillance-system. The denominator was the true total HIV cases arising from a Negative Binomial distribution. The trial probability, known as ascertainment probability, depended on socio-economic determinants of HIV, retained via feature-selection algorithms. Data were obtained from CDC's HIV report for End of the HIV Epidemic and the American Community Survey. The prediction model was assessed out-of-sample in Georgia counties. We combined socio-economic data with the posterior predictive distribution of the coefficients to predict the mean ascertainment probability and total HIV cases at the ZIP Code-level. These estimates were spatially smoothed and aggregated at the county-level for secondary validations.

Results

The county-level model showed good mixing properties and predictive accuracy. The mean ascertainment probability calibrated to the ZIP Code-level varied from 78.4% (95% credible interval: 24.4%–99.3%) to 93.8% (95%CI: 80.6%–99.8%). Further, the predicted undiagnosed HIV cases ranged between 12 (95%CI: 6–19; ZIP Code 30322) to 1603 (95%CI 1209–1968; ZIP Code 30318).

Conclusions

Our findings provide a more complete picture of the relative burden of HIV across ZIP codes. Such information can be used by Local Health Departments to identify underserved areas and allocate resources accordingly. Furthermore, our methodological approach can be applied to complement the information obtained from passive surveillance, especially when more resource-intensive approaches are not available or are unfeasible to employ.